In Todayโs Issue:
๐ OpenAI's first device wants to live in your house
๐ Anthropic makes Claude free for every US teacher
โ๏ธ Hassabis asks for a referee, the White House says no
๐ Codex adds a million users a day
๐ Publishers drag Google's training data into court
โจ And more AI goodnessโฆ
โก The Signal
Every big story today is about a limit: how much compute OpenAI can buy, how much of your home it can occupy, and how much of the industry Demis Hassabis thinks should be policed by someone other than itself.
Codex climbed from six million users to eight million in forty-eight hours this week, three days after GPT-5.6 Sol shipped, and OpenAI's own Codex lead spent this morning publicly asking users whether he should reset their limits again, while Sam Altman warned that things may start to break. In the same week, OpenAI let it be known that its first machine is a camera-equipped speaker built to sit in your kitchen and learn you from your email. Hassabis wants a FINRA-style body testing frontier models before release, and a White House adviser has already told him there will be no FDA for AI. Anthropic, meanwhile, is taking the least glamorous route into the same future, one verified teacher at a time. The pattern underneath all of it is that capability stopped being the hard part a while ago, and capacity, courtrooms and consent are what actually ration this technology now.
All the best,
Kim Isenberg
(Claude for Teachers building a Grade 7 maths lesson. Anthropic)
๐ Anthropic Puts Claude in Front of Every US Teacher, Free
Anthropic is giving every verified US K-12 teacher free access to premium Claude, and it built the product with the American Federation of Teachers looking over its shoulder. Claude for Teachers, launched Tuesday, covers lesson planning, differentiated instruction, formative assessments and parent communications, and adds a Learning Commons connector that pulls academic standards from all 50 states plus hooks into classroom tools such as MagicSchool, Canva Education and Diffit. Model training is off by default, the terms are written to comply with FERPA, and the package normally costs $20 a month; teachers who verify by June 30, 2027 get a full year free.
๐ tl;dr: Free premium Claude for US teachers, with state standards built in and student data walled off.
(Demis Hassabis. TechCrunch / Getty Images)
โ๏ธ DeepMind's CEO Wants a Referee for Frontier AI
Demis Hassabis wants an independent standards body for frontier AI, modelled on FINRA, the industry-funded organisation that polices Wall Street brokers. Writing on X on Tuesday, the DeepMind chief proposed that labs begin by voluntarily sharing models for review up to 30 days before release, with the arrangement turning mandatory for US deployment once it proves itself; the body would test models, set release practices and handle post-release vulnerabilities, staffed by technical experts and open-source representatives and funded by the labs. "The strength of this approach is it would be technically focused, while at the same time supporting innovation," he wrote. White House AI adviser Sriram Krishnan has already pushed back, saying "there will not be an FDA for AI."
๐ tl;dr: The head of a lab that would be policed is asking for a policeman; the White House says no.
(Sam Altman. TechCrunch / Getty Images)
โ ๏ธ Altman Warns the Servers May Buckle
Sam Altman told users to expect "hiccups" as demand for GPT-5.6 Sol outruns OpenAI's ability to serve it. "5.6 sol growth is insane. the inference team has done heroic work to be able to support demand. we are going to move mountains to continue to scale, but it is possible there are some hiccups soon," he wrote on X on Tuesday, five days after the model went public. It is a useful corrective to the idea that frontier AI is rate-limited by clever ideas; OpenAI's constraint this week is chips, electricity and queueing.
๐ tl;dr: OpenAI's flagship is growing faster than OpenAI can serve it.
Turn a chatbot into a tutor for the one thing you keep pretending to understand.
Why it helps: The teaching tools Anthropic shipped today work because they force the boring part up front: the goal, the level, and the misconception to watch for. Any chatbot will do the same if you make it, and most people never ask.
Try this: "I want to understand [topic] well enough to explain it to a sharp colleague in five minutes. First, ask me three questions to work out what I already know and where I am likely to be wrong. Then teach me in stages, and after each stage stop and test me with one question. Do not move on until I answer it correctly."
๐ฌ Watch This
Rich Sutton thinks the industry that built today's AI took a wrong turn, and he brought an architecture to prove it.
Sutton co-wrote the founding textbook of reinforcement learning and shared the 2024 Turing Award for that work, which makes his opening claim in this RLC 2025 talk land harder than it would from anyone else: AI has become a huge industry and has largely lost its way. His objection is that today's models are trained once and then frozen, while anything worth calling intelligence keeps learning from its own experience.
His answer is OaK, a design in which every component keeps learning, each individual weight tunes its own learning rate as it goes, and the agent builds its own abstractions by noticing a useful feature, inventing a subtask around it, learning to solve that subtask, modelling the result, and planning with it. It is a direct challenge to the pretrain-then-ship orthodoxy that every other model in today's issue is built on.
Most companies donโt have an AI access problem. They have an execution problem.
I keep hearing the same pattern: Claude is in employeesโ hands, people are moving faster, but the business itself hasnโt changed. The valuable work is still trapped in individual chats and isolated experiments.
That gap helps explain why Forward Deployed Engineers have become one of the most talked-about deployment models in AI. Rather than advising from the outside, FDEs embed with teams and turn isolated AI use into enterprise AI solutions.
Itโs also what made me pay attention to HatchWorks AI. As an official Anthropic partner, HatchWorks AI embeds Anthropic-certified FDEs to identify a high-value business problem, build and deploy the solution, put governance around it, and train the team to keep improving it.
If your Claude rollout is still mostly individual usage, their approach is worth a serious look.
"Embarrassment of riches. But looks like we might hit 9M soon. Should we reset the ChatGPT Work and Codex usage again or give it some space?"
โ Tibo Sottiaux, Codex and ChatGPT at OpenAI, on X
The man running Codex is asking his own users whether he should hand out more capacity. Codex and ChatGPT Work climbed from six million to eight million active users in forty-eight hours this week, and OpenAI has been resetting usage limits almost daily to keep pace. Today's Graph of the Day is that curve.
๐ On the same morning DeepMind's chief executive called for an independent referee for frontier AI, Google's own training data landed in a Manhattan courtroom.
(The Gemini app, the product at the centre of the complaint. TechCrunch / Getty Images)
Hachette, Cengage, Elsevier and the novelist Scott Turow have filed a proposed class action in the Southern District of New York alleging that Google built Gemini on millions of copyrighted books and journal articles, including works the publishers had handed over strictly for Google Books search. The complaint claims Google "illegally copied countless books and journal articles obtained for strictly limited use" and "downloaded unauthorized web scrapes of virtually the entire internet", and it points to an internal Google assessment that reportedly warned of "$10Bs-$100Bs" in potential fines.
Google has not responded publicly to the specific allegations; the plaintiffs are seeking damages and an injunction.
OpenAI's First Computer Wants to Live With You
The Takeaway
๐ OpenAI's debut device is a screenless smart speaker with a camera, sensors, a rechargeable battery and mechanical parts that move on their own, according to Bloomberg.
๐ It runs on GPT-Live, the listen-and-talk-at-once voice model from this month, and is built to feel like a companion that learns you over time.
๐ The hardware group is largely ex-Apple: Jony Ive's io ($6.5B), LoveFrom, Tang Tan, Evans Hankey, Paul Meade and 400+ Apple hires.
๐ OpenAI wants to unveil it this year and ship in 2027, but Apple is seeking an injunction that could delay sales.
OpenAI's first piece of hardware asks for something no chatbot has ever asked for: a permanent place in your home.
Bloomberg, whose Mark Gurman first reported the design on Tuesday, describes a screenless smart speaker that can move on its own. It carries a camera and sensors to read the room, a rechargeable battery so it travels from the kitchen to the bedroom, and mechanical parts that shift by themselves to suggest something alive rather than an appliance waiting for orders. It runs on GPT-Live, the voice model OpenAI shipped this month that can listen and talk at the same time. Internally, OpenAI calls it the first computer built for AI, and it is meant to grow more personal over time, drawing on material as intimate as your emails to anticipate what you need before you ask. Sonos shares fell more than 10% in late trading on the report; Apple slipped less than 1%.
(Sam Altman, whose hardware bet is now in court. TechCrunch / Getty Images)
The team building it is largely Apple's. OpenAI paid $6.5 billion last year for io Products, co-founded by former Apple design chief Jony Ive, whose studio LoveFrom is shaping the lineup. Tang Tan, who ran iPhone product design, is now OpenAI's Chief Hardware Officer. Evans Hankey, Apple's former head of industrial design, leads the speaker. Last month OpenAI hired Paul Meade, who built the Vision Pro. Apple's suit alleges the pipeline carried more than people: it accuses Tan of leading a campaign to obtain confidential information about Apple's future products, and counts more than 400 former Apple employees now at OpenAI. OpenAI says it has "no interest in other companies' trade secrets" and that it is "not aware of any evidence that this complaint has merit."
The sharper question is the one OpenAI is asking of its customers. A machine that watches a room, listens continuously and reads your mail is both a remarkable assistant and a remarkable collection surface, and OpenAI has not said how much of what it sees and hears will leave the house. Legally, the company argues that its audio system and hardware differ substantially from Apple's HomePod and that its first product infringes nothing. Apple is building a rival family regardless, beginning with J490, a long-delayed smart-home hub with a square 7-inch display, videoconferencing and facial recognition, built to showcase the new Siri in iOS 27.
(Apple's HomePod, which OpenAI says its own speaker will not resemble. Photographer: Mark Kauzlarich / Bloomberg)
OpenAI has roughly five products in development and plans to unveil the speaker this year for a 2027 release, with a smartphone replacement, a wearable pendant and home robotics behind it. Apple is seeking an injunction that could stall the lot. The company that taught the world to type into a text box now wants to be the thing you talk to across the room, and two obstacles stand between it and your kitchen counter: a courtroom, and your own comfort with a camera in the room.
Why it matters: The smartphone made you reach for the assistant; a proactive speaker with a camera, a battery and access to your inbox reverses that, and the entry price is letting OpenAI's models see and hear your home. Apple's injunction is the nearest threat to the timeline, but the privacy question is what decides whether anyone puts one on the counter.
Sources:
๐ https://www.bloomberg.com/news/articles/2026-07-14/openai-says-it-s-not-aware-of-any-evidence-that-apple-lawsuit-has-merit
๐ OpenAI's first hardware device is reportedly a screenless speaker that can move (TechCrunch)
๐ OpenAI's First Device Will Be Movable, Screenless Speaker Built as AI Companion (Bloomberg)
The chart: A reader-built timeline of OpenAI Codex active users through 2026, assembled from Tibo Sottiaux's own milestone posts. Codex sat at 3M on April 9 and took almost two months to reach 5M on May 31. Then GPT-5.6 Sol shipped on July 9 and the line goes near vertical: 6M on July 12, 7M on July 13, 8M on July 14. The dashed run up to 9M is a projection, not a milestone anyone has reached.
The lesson: The adoption cliff tracks the model release rather than any marketing push. Codex needed seven weeks to add its fourth and fifth million, from April 9 to May 31, and forty-eight hours to add its seventh and eighth, from July 12 to July 14, with the inflection landing three days after Sol went public. That same curve is why Sam Altman spent Tuesday warning about "hiccups": a line this steep is an inference bill before it is a victory lap.
The caveat: The chart is labelled Codex, but Sottiaux's underlying posts count active users "across Codex and ChatGPT Work", so the line quietly bundles two products. Active user is never defined, the milestones are announcements rather than measured telemetry, and the tidy one-million-a-day steps look like rounded PR beats rather than a real curve. Trust the shape, not the precision.
๐ง The Open Model That Sees and Hears Without a Translator
โก Bottom line
Google's Gemma 4 report explains how an open model family of 2.3B to 31B parameters learned to read raw images and audio.
๐ก Why it matters
Capable multimodal AI now runs on hardware you own, with nothing leaving the room to reach a cloud.
๐ What it means
The distance between frontier labs and models anyone can download keeps shrinking, faster than the labs would like.
Most AI that "sees" works like a tourist with a translator. The picture goes to a separate vision encoder, which converts it into something the language model can read, and the language model never touches the original. Google's Gemma 4 technical report, posted to arXiv on July 2, describes what happens when you fire the translator.
The family runs from 2.3B to 31B parameters and mixes two designs: dense models, where the entire network fires for every word, and Mixture-of-Experts, where only a fraction of it wakes up for any given word, which is how a large model stays cheap to run. The interesting choice is the 12B model, which drops the dedicated encoders and takes raw image patches and raw audio straight in. The paper's own audio results make the case: competitive audio performance, it argues, is achievable with no dedicated audio encoder at all.
(How Gemma 4 carves a picture into patches before the model reads it. Figure 2, Gemma 4 Technical Report)
Two further choices matter to anyone who actually runs these models. Gemma 4 adds a thinking mode, so it can write out its reasoning before committing to an answer, and a multi-token prediction drafter that sketches several words ahead instead of strictly one at a time, which is a large part of why it feels quick.
(The multi-token prediction drafter, fed by the main model. Figure 1, Gemma 4 Technical Report)
Google reports gains over Gemma 3 across STEM, multimodal and long-context tests, and on Arena Text, where humans compare answers blind without knowing which model wrote them, Gemma holds its own against open models several times its size. The point for a curious reader is smaller and sharper than any benchmark table: what recently needed a data centre now fits on a workstation, and it listens.
Slack replies in seconds. Not minutes.
Dictate into Slack, email, LinkedIn, or any app and get polished, send-ready text. Wispr Flow strips filler and formats everything. 89% of messages sent with zero edits. Works on Mac, Windows, and iPhone.
